Infrared target recognition with deep learning algorithms
Laixiang Xu,Fengjie Zhao,Peng Xu,Bingxu Cao
DOI: https://doi.org/10.1007/s11042-022-14142-x
IF: 2.577
2022-11-13
Multimedia Tools and Applications
Abstract:Infrared automatic target recognition (ATR) technology still is a challenging problem in military applications. In recent years, convolutional neural networks (CNNs) models have already led to breakthrough developments in object detection and target recognition. However, the complex environment and the bad weather caused the poor texture information and the weak background of infrared imaging. It's difficult to use standard CNNs to perform accurate feature extraction and target classification. To overcome these shortcomings, we propose a novel deep learning framework, composed of the multi-kernel transformation and the Alpha-Beta divergence. The multi-kernel transformation operation is designed between convolutional layers and pooling layers to increase the confidence of feature extraction. The Alpha-Beta divergence is used as a penalty term to re-encode the output neurons of improved CNNs, which can promote the recognition performance of the entire network. Furthermore, comprehensive theoretical analysis and extensive experiments are confirmed that our proposed framework outperforms ResNet, VGG-19, DenseNet, and the different combinations of models in many aspects, such as short time-consuming, high accuracy, and strong robustness. Our approach yields a maximum accuracy score of 98.43% on our dataset. Meanwhile, we use the OKTAL-SE-based synthetic database and the SENSIAC dataset to verify our models. Experimental results demonstrate the maximum average accuracy is 97.16%, it is feasible and effective for infrared target recognition.
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering